metadata
language:
- en
license: apache-2.0
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- trl
- sft
- code
- lora
- peft
base_model: unsloth/tinyllama-chat-bnb-4bit
pipeline_tag: text-generation
datasets: Ramikan-BR/data-oss_instruct-decontaminated_python.jsonl
Uploaded model
- Developed by: Ramikan-BR
- Model type: [text-generation/Python Coder]
- Language(s) (NLP): [en]
- License: apache-2.0
- Finetuned from model : unsloth/tinyllama-chat-bnb-4bit
Model Description
Training Data
datasets: Ramikan-BR/data-oss_instruct-decontaminated_python.jsonl
Training Procedure
The model was refined using Unsloath. The dataset ise-uiuc/Magicoder-OSS-Instruct-75K was adjusted, leaving only data on python and divided into 10 parts, each refinement occurred for 2 epochs, using adafactor optimizer or adamw_8bit (adafactor seems to deliver less loss).
Model Sources [optional]
base_model: unsloth/tinyllama-chat-bnb-4bit
model: Ramikan-BR/tinyllama-coder-py-4bit-v10 gguf_f16: tinyllama-coder-py-4bit-v10-unsloth.F16.gguf gguf_Q4_K_M: tinyllama-coder-py-4bit-v10-unsloth.Q4_K_M.gguf gguf_Q8_0: tinyllama-coder-py-4bit-v10-unsloth.Q8_0.gguf
Training Hyperparameters
Notebook Unsloath that I used for AI refinement: TinyLlama
%%capture
# Installs Unsloth, Xformers (Flash Attention) and all other packages!
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers trl peft accelerate bitsandbytes # xformers "xformers<0.0.26"
import os
from google.colab import drive
drive.mount('/content/drive')
from unsloth import FastLanguageModel
import torch
max_seq_length = 4096 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
# 4bit pre quantized models we support for 4x faster downloading + no OOMs.
fourbit_models = [
"unsloth/mistral-7b-bnb-4bit",
"unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
"unsloth/llama-2-7b-bnb-4bit",
"unsloth/llama-2-13b-bnb-4bit",
"unsloth/codellama-34b-bnb-4bit",
"unsloth/tinyllama-bnb-4bit",
"unsloth/gemma-7b-bnb-4bit", # New Google 6 trillion tokens model 2.5x faster!
"unsloth/gemma-2b-bnb-4bit",
] # More models at https://huggingface.co/unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "Ramikan-BR/tinyllama-coder-py-4bit_LORA-v9", # "unsloth/tinyllama" for 16bit loading
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
model = FastLanguageModel.get_peft_model(
model,
r = 256, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",],
lora_alpha = 512,
lora_dropout = 0, # Currently only supports dropout = 0
bias = "none", # Currently only supports bias = "none"
use_gradient_checkpointing = True, # @@@ IF YOU GET OUT OF MEMORY - set to True @@@
random_state = 3407,
use_rslora = False, # We support rank stabilized LoRA
loftq_config = None, # And LoftQ
)
alpaca_prompt = """Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
{}
### Output:
{}"""
EOS_TOKEN = tokenizer.eos_token
def formatting_prompts_func(examples):
inputs = examples["problem"]
outputs = examples["solution"]
texts = []
for input, output in zip(inputs, outputs):
# Must add EOS_TOKEN, otherwise your generation will go on forever!
text = alpaca_prompt.format(input, output) + EOS_TOKEN
texts.append(text)
return { "text" : texts}
pass
from datasets import load_dataset
dataset = load_dataset('json', data_files='/content/drive/MyDrive/data-oss_instruct-py-10.jsonl', split='train')
dataset = dataset.map(formatting_prompts_func, batched=True)
from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported
from transformers.utils import logging
logging.set_verbosity_info()
trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
dataset_text_field = "text",
max_seq_length = max_seq_length,
dataset_num_proc = 2,
packing = True, # Packs short sequences together to save time!
args = TrainingArguments(
per_device_train_batch_size = 2,
gradient_accumulation_steps = 256,
warmup_ratio = 0.1,
num_train_epochs = 2,
learning_rate = 2e-4,
fp16 = not torch.cuda.is_bf16_supported(),
bf16 = torch.cuda.is_bf16_supported(),
logging_steps = 1,
optim = "adafactor", # adamw_torch ou adamw_torch_fused +10% velocidade ou adafactor ou adamw_8bit
weight_decay = 0.1,
lr_scheduler_type = "linear",
seed = 3407,
output_dir = "outputs",
),
)
trainer_stats = trainer.train()
model.save_pretrained("lora_model") # Local saving
tokenizer.save_pretrained("lora_model")
model.push_to_hub("Ramikan-BR/tinyllama-coder-py-4bit_LORA-v10", token = "hf_...") # Online saving
tokenizer.push_to_hub("Ramikan-BR/tinyllama-coder-py-4bit_LORA-v10", token = "hf_...") # Online saving
# Merge to 16bit
model.save_pretrained_merged("model", tokenizer, save_method = "merged_16bit",)
model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "merged_16bit", token = "hf_...")
# Merge to 4bit
if False: model.save_pretrained_merged("model", tokenizer, save_method = "merged_4bit",)
if False: model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "merged_4bit", token = "hf_...")
# Just LoRA adapters
if False: model.save_pretrained_merged("model", tokenizer, save_method = "lora",)
if False: model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "lora", token = "hf_...")
# Save to 8bit Q8_0
model.save_pretrained_gguf("model", tokenizer,)
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, token = "hf_...")
# Save to 16bit GGUF
model.save_pretrained_gguf("model", tokenizer, quantization_method = "f16")
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, quantization_method = "f16", token = "hf_...")
# Save to q4_k_m GGUF
model.save_pretrained_gguf("model", tokenizer, quantization_method = "q4_k_m")
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, quantization_method = "q4_k_m", token = "hf_...")
Loss for 5 epochs in the last training session of the last part of the dataset:
==((====))== Unsloth - 2x faster free finetuning | Num GPUs = 1
\\ /| Num examples = 407 | Num Epochs = 5
O^O/ \_/ \ Batch size per device = 2 | Gradient Accumulation steps = 256
\ / Total batch size = 512 | Total steps = 5
"-____-" Number of trainable parameters = 201,850,880
[5/5 29:36, Epoch 3/5]
Step Training Loss
1 0.568000
2 0.145300
3 0.506100
4 0.331900
5 0.276100
Quick test 1 after training the last part of the dataset:
# alpaca_prompt = Copied from above
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
alpaca_prompt.format(
"Continue the fibonnaci sequence.", # instruction
"1, 1, 2, 3, 5, 8", # input
"", # output - leave this blank for generation!
)
], return_tensors = "pt").to("cuda")
AI Response: ['<s> Below is an instruction that describes a task. Write a response that appropriately completes the request.\n### Input:\nContinue the fibonnaci sequence.\n\n### Output:\n1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 420, 787, 1444, 2881, 4765, 8640']
Quick test 2 after training the last part of the dataset:
# alpaca_prompt = Copied from above
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
alpaca_prompt.format(
"Continue the fibonnaci sequence.", # instruction
"1, 1, 2, 3, 5, 8", # input
"", # output - leave this blank for generation!
)
], return_tensors = "pt").to("cuda")
from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)
AI Response: <s> Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
Continue the fibonnaci sequence.
### Output:
1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 420, 787, 1444, 2881, 4765, 8640, 17281, 31362, 65325, 128672, 251345, 410000, 720000, 1280000,
Quick test 3 after training the last part of the dataset:
if False:
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "lora_model", # YOUR MODEL YOU USED FOR TRAINING
max_seq_length = max_seq_length,
dtype = dtype,
load_in_4bit = load_in_4bit,
)
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
# alpaca_prompt = You MUST copy from above!
inputs = tokenizer(
[
alpaca_prompt.format(
"What is a famous tall tower in Paris?", # instruction
"", # input
"", # output - leave this blank for generation!
)
], return_tensors = "pt").to("cuda")
from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 64)
AI Response: <s> Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
What is a famous tall tower in Paris?
### Output:
The famous tall tower in Paris is the Eiffel Tower. It is a 300-meter-tall steel tower located in the heart of Paris, France. The tower was built in 18892 and is a popular tourist attraction. It is also a symbol of the city
outputs = model.generate(**inputs, max_new_tokens = 64, use_cache = True)
tokenizer.batch_decode(outputs)
This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.
